A unified Cross-Modal adaptive contrastive learning framework for acoustic fault diagnosis of rolling bearing under limited & imbalanced data
Linhao Peng, Fang Liu, Ang Lu, Yongbin Liu, Changqing Shen, Min Xia
Abstract
The intelligent fault diagnosis method based on acoustics provides a effective approach for achieving reliable data-driven maintenance in industrial scenarios. However, in actual scenarios, mechanical equipment is mainly in a normal operating state, which leads to limited monitoring data and an imbalance in fault categories (L&I). Moreover, multi-source heterogeneous data are often fragmented and difficult to integrate, while most artificial intelligence models merely provide fault warnings without dynamic optimization or decision-making capabilities. To address these issues, this paper proposes a unified cross-modal adaptive contrastive learning framework jointly models the ideas of data-level augmentation and algorithm-level adaptation within a single end-to-end architecture. From a data-level perspective, inspired by the multi-view representation theory, the proposed method constructs a same-source dual-modality input by transforming raw acoustic signals into temporal waveforms and gramian angular difference field (GADF) images, thereby enriching feature diversity without introducing sensor inconsistency. A multi-scale residual image encoder (MSR-IE) and a hybrid temporal encoder with multi-receptive convolutions (HTE-MRC) are designed to extract modality-specific features. Subsequently, integration is carried out through a cross-modal contrastive fusion module. During this process, heterogeneous feature Spaces are aligned into a unified, domain-invariant embedding. At the algorithmic level, a dual-task guided fault discriminator (DGFD) is developed to jointly perform coarse-grained health assessment and fine-grained fault identification, with dynamic task reweighting to balance learning under class-imbalanced conditions. Results on two rolling bearing acoustic datasets across six imbalance regimes show that the proposed method achieves 97.9 % accuracy and 98.9 % G-mean, with notably improved minority-class detection and balanced performance under severe imbalance.